Abstract
Conventional building control and automation relies on a central decision logic that controls the in-building environment based on readings aggregated from multiple sensors distributed throughout the building. The operating mode is detect-then-decide, such that building control decisions are made only after data is received from the suite of sensors. These techniques do not work well when there are missing observations and are therefore energy intensive. Further, the central decision logic needs to be adapted to building-specific factors. This disclosure describes the use of a large language model (LLM) to automatically identify sensors in a building that can be turned off without affecting building control tasks. Per the techniques, raw sensor data from sensors in the building along with metadata such as sensor location, time-of-day, etc. are utilized to form text snapshots. The text snapshots are tokenized to obtain feature vectors. The feature vectors and a custom prompt specifying the task are provided as input to an LLM. The LLM output response is parsed to identify sensors that can be turned off. The prompt can be customized as necessary. The LLM reasoning capability is relied upon to interpret sensor observations in the context of sensor locations and other data to identify sections of sensors that can be turned off.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Shin, D, "Building Control and Sensor Power Use Optimization Using a Large Language Model", Technical Disclosure Commons, (December 21, 2023)
https://www.tdcommons.org/dpubs_series/6531